Image compression using optimal transport mapping based on ranking visual saliency

Due to the growth of multimedia technology and increased popularity of virtual reality (VR) and augmented reality (AR) applications, the need for efficient image transmission and storage is also proliferating. We propose an image compression method based on image region importance recognition and optimal transport (OT). As not all regions in an image is equally important, we can identify critical regions in the image for high-fidelity compression and use low-fidelity compression for less critical regions. In compression, critical regions are enlarged while less important regions are reduced in size. In this process we use OT to assign higher importance to the critical regions. In uncompression, inverse OT is used to restore the image. Our method has the advantages that, first, it is applicable to various compression frameworks, including both traditional and deep learning-based compression frameworks, and second, it is more flexible in adjusting the foreground and background of the image, and the ratio of the foreground to the background can be freely adjusted according to the actual needs of application. Experimental results show that our method can improve the quality of restored image while achieving a high compression rate. Comparisons under varying experimental conditions show that our method outperforms existing techniques at the same level of compression rate by improving reconstruction fidelity for key regions by 4.5–7.0 % (95 % confidence interval).